Warning, this model has discovered NSFW and can randomly produce a NSFW image when not expecting one.
58 block layers
12 runs per single block layer to find best weights 0.3 to 1.4
696 total weights tested block by block to get 58 weights for the merge
So something like this, pick the best pic from each block layer
Get all the block weights hand picked, merge the model with the new weights. Isn't there a lot of variables to this? Sure, but I kept it close to my normal gens and after testing, it seems fine with other samplers, schedulers, loras, etc..
A much improved version with a constant ability to reproduce better consistant images. Q6 and Q4 models produce much better images out of the gate now. Does NSFW, but with the typical flux anatomy issues, if you struggle with flux not wanting to do NSFW try adding "exposed (breast:2.0), (Nipples:2.0)" You will need the Vae and Dual Clip as these are unet models.
Not of a huge fan of Schnell on first use so I took up a challenge to remove what I didn't like of Schnell. First goal was to remove the hideous Greenish blotchy look and the random dark splotches all over. I used lora and thanks to the negative field was able to remove most of those issues and get a more natural looking image. So I am calling this version a win. The standard model produces much better images out of the gate vs the ggufs. A refiner is definately needed in both cases, the ggufs are complete rubish looking straight out of the sampler, but it's a good thing as the noise gives the refiner something good to work with. So find a favorite refiner model and plug it in for best results. Images are created in 8 steps. I only tested the 8_0K gguf a couple times, not sure how well the K6 and K4 models do. I do like the artsy style of schnell tho, it can really produce some unique beatuiful pictres in a hurry.
Standard Workflow
GGUF Workflow
Standard Comparison sampler vs refine
Standard
GGUF Q8
Standard with controlnet
Link below are more images - the latest, most upper ones
